An IoT Based Smart Farming System Using
Machine Learning
A.Dahane 1,2,3, R.Benameur 2,3
1Institute
of Applied Science and Technology ISTA
2Research Laboratory in Industrial Computing and Networks (RIIR)
3University of Oran 1, Algeria
E-mail: dahane.amine@univ-oran1.dz
Abstract— Smart farming allows to analyze the growth of
plants and to influence the parameters of our system in real time
in order to optimize plant growth and support the farmer in his
activity. Internet of Things (IoT) arrangements, based on the
application particular sensors data measurements and intelligent
processing, are bridging the holes between the cyber and physical
worlds. In this paper, we propose the design and the experiment
of a smart farming system based on an intelligent platform which
enables prediction capabilities using artificial intelligence (AI)
techniques. This system is based on the technology of wireless
sensor networks and its implementation requires three main
phases, i) data collection phase using sensors deployed in an
agricultural field, ii) data cleaning and storage phase, and iii)
predictive processing using some AI methods.
Keywords—Smart farming; IoT; Machine learning; Sensing;
Precision agriculture.
I. INTRODUCTION
Agriculture uses 85% of available freshwater resources
worldwide, and this percentage will continue to be dominant
in water consumption because of population growth and
increased food demand. There is an urgent need to develop
strategies based on the science and the technology [1]. The
need of using IoT in agriculture applications becomes
nowadays a reality after the success of theoretical research
contributions in the previous decade [2,6]. IoT have recently
motivated adoption of information technology solutions in
crop fields within precision agriculture approaches which
contribute for rational use of water, including technical,
agronomic, managerial,…etc. Several research studies were
conducted on irrigation systems to achieve water savings in
diverse crops from basic ones to more technologically
advanced ones [4,5,7,8,9 and 21]. To achieve water saving,
irrigation system frameworks have been proposed based on
various techniques, e.g., thermal imaging, Crop Water Stress
Index (CWSI), direct soil water measurements, etc. Precision
farming is a principle that aims to optimize yields and
investments, while seeking to better take into account the
variability of the environment. This concept appeared in the
context of the race for progress in agricultural yields. It
notably influenced: tillage, sowing, fertilization, irrigation,
spraying of pesticides. Precision agriculture is essential to
meet the challenges of agricultural production in terms of
productivity, environmental impact, food security and
978-1-7281-5628-6/20/$31.00 ©2020 IEEE
Authorized licensed use limited to: Cornell University Library. Downloaded on May 24,2021 at 15:31:06 UTC from IEEE Xplore. Restrictions apply.
B.Kechar 2,3, A.Benyamina 3,4
2Research Laboratory in Industrial Computing and Networks (RIIR)
3University of Oran 1, Algeria
4LAPECI Laboratory, University of Oran1
E-mail: kechar.bouabdellah@univ-oran1.dz
sustainability as the world's population demands a significant
increase in food production must be achieved while now
availability and high nutritional quality worldwide [14,16 and
18]. Due to the recent advances in sensor technology for the
implementation of irrigation systems in agriculture especially
smart farming [13, 15 and 17] and the evolution of IoT
technologies that can be applied in the development of these
systems. IoT takes a major role in the revolutions of digital
transformations, and today we will discuss the applications of
‘IoT in Agriculture’. We anticipate the proposal of IoT-based
intelligent irrigation architecture with a machine learning
approach to predict soil moisture to reach more convincing
conclusions. In this paper, we propose an IoT based smart
farming architecture founded on new EDGE-Fog-IoT-Cloud
platform. After giving the design of the overall architecture,
we detail the implementation part (hardware and software
components). The aim of this platform is to demonstrate the
effectiveness of AI techniques to help in making effective
irrigation decisions with optimum water usage in smart
farming. The rest of the paper is organized as follows. Section
II briefly surveys the related works. Section III presents our
proposed smart farming platform in agriculture of precision
specially irrigation. Section VI introduces and explains the
tool developed and illustrates the use of AI techniques.
Section V concludes the paper and outline directions of future
work.
II. RELATED WORKS
We present in this section recent research work aimed at the
implementation of irrigation systems in agriculture. Some of
them have used AI techniques to enhance the prediction
aspect. Researchers in [3] developed irrigation sensor based on
smart phone. For sensing soil moisture, the digital camera of
smart phone is used to process RGB to gray for estimation of
ratio between wet and dry area of soil. The ratio of wetness
and dryness is transmitted via gateway to water motor
controller. A Mobile Application (APP) is developed to
control sensor activity (like wakeup) and to set sensor in sleep
mode. Majority of the earlier irrigation systems do not
consider the weather forecasting information (e.g.,
precipitation) while making irrigation decisions. It leads to a
wastage of fresh water, energy and loss of crop growth (due to
excess water) when a rain is followed immediately by the
watering of the crop.
To handle such cases, WSNs based solutions can provide a
better decision support for irrigation by utilizing weather
forecasting information (e.g., precipitation) from the Network.
Irrigated agriculture represents the bulk of the demand for
water, being the first sector affected by water shortage,
resulting in a decreased capacity to maintain per capita food
production. Therefore, the efficient use of water in agriculture
is one of the most important agricultural challenges that
modern technologies are helping to resolve [11]. The objective
of authors in [10] was to develop understanding on site-
specific suitability of the Mid Elevation Spray Application
(MESA) and Low Elevation Spray Application (LESA)
sprinkler systems in irrigating corn crop and potential water as
well as energy savings. The automated irrigation system put in
place has proven to be feasible and cost-effective to optimize
water resources for agricultural production [1,5 and 21].
Authors in [8] propose an intelligent irrigation architecture
based on IoT as well as a hybrid approach based on deep
learning to predict soil moisture. The proposed algorithm uses
sensor data from the recent past and weather forecasts to
predict soil moisture for the next few days. The disadvantage
of these approaches [1,5] is that they set the maximum
humidity and do not take into account the need for water
exactly needed. They plan to save water. Unlike this state-of-
the-art work, we propose in this paper an automated irrigation
management platform using a WSN by making the following
contributions: i) Design and implementation of a new EDGE-
Fog-IoT-Cloud based architecture dedicated to the smart
farming, and ii) We show clearly the interest of our system in
saving water and preserving this natural resource by using AI
techniques to predict soil moisture of the upcoming days.
III. SMART FARMING SYSTEM
The smart farming platform architecture (Fig. 2, Fig. 3) has
been proposed to collect, transmit and process the physical
parameters (Soil-Moisture, Air-Temperature, Air-Humidity,
Water-level, Water-flow, Luminous intensity, Combustible
Gas (vapors) for security of crops) of farming land along with
Fig. 1: Hardware components involved on the Field experiment. [Legends 1: DHT22 Sensor, 2: LDR Sensor, 3: Water level Sensor, 4: Relay Switch, 5: Water
pump, 6: Power supply 9 V, 7: Water flow sensor, 8: Soil moisture sensor, 9: LCD display, 10: MQ2 sensor, 11: LEDs, 12: Jumpers]. The mega arduino card
and NRF module are in the Box.
Authorized licensed use limited to: Cornell University Library. Downloaded on May 24,2021 at 15:31:06 UTC from IEEE Xplore. Restrictions apply.
the weather forecast information to manage the irrigation
efficiently. The system architecture is designed in three layers:
data layer which is depicted in Fig. 1, data processing layer
and application layer. This distributed architecture is shown in
Fig. 2 and Fig. 3 which represents a major extension of our
previous works [20,21] on agriculture of precision where we
calculated the useful reserve of water (RU) according to the
texture of the soil (Clay, Limon, Sand, Organic Matter), and
for the second method we used the (Rawls and Turq
formulas) equations. The operational architecture shows
information regarding soil moisture of the upcoming days, it
also provides irrigation suggestions, based on the defined level
of soil moisture and predicted precipitation, to save water and
energy the generated information by algorithm and device is
stored in MySQL database at the server (see Fig. 2). A web
service based on NodeJS process data before sending them
periodically to the cloud architecture. All streams of data are
stored in Google drive before they are transmitted to Google
Colab to execute artificial intelligence techniques (AI) used
for training, testing and prediction. The EDGE is placed
between IoT sensors Network and the Cloud as an
intermediate layer. EDGE comprises three primary modules:
data acquisition, prediction, and visualization. Data sent by
sensors network by means of different transmission
network protocols. Data is stored closely to the EDGE to
provide quick access to data for farmer during interventions.
Moreover, farming data remains available even if the internet
connection is temporarily interrupted. Processing and
validating data at EDGE level reduce the amount of data
transmitted to the cloud and conserving global energy of the
architecture and network bandwidth. Each sensor in our IoT
system has a range of measurement which must be evaluated
in the level of EDGE to serve as a baseline for the detection of
abnormal behavior. For example, we have temperature
measuring range -40 to +80 °C for the DHT22 sensor. These
not clean data must also be corrected according to the
behavior of each sensor and also take into account historical
data to avoid triggering unnecessary alerts and adapt the
frequency of collecting sensing data in order to obtain clean
data.
Fig. 2: Global View of our proposed operational architecture.

Fig. 3: Proposed layered architecture of smart farming with the technical choices of implementation.

Fig. 4: Main steps in the proposed architecture based on IA techniques.

Authorized licensed use limited to: Cornell University Library. Downloaded on May 24,2021 at 15:31:06 UTC from IEEE Xplore. Restrictions apply.
 The designed platform is based on the ideas proposed by
[1, 5, 8 and 19] with modifications made for our application.
As we can see in Fig. 2, Fig. 3, the devices of our platform in
the first layer aim to collect field data and transmit them using
NRFL01 radio module. These data can be stored in private
database so that the user (farmer) has a possibility to analyze
them and monitor the crops in real time. In order to start
irrigation planning we need to calculate the UR parameter in
mm.
IV. EXPERIMENTATION
This section presents the implementation of the new
EDGE-Fog-IoT-Cloud based architecture dedicated to smart
farming and the analysis of the obtained results.
A. Implementation and results analysis
The main objective of these experiments is to show how to
create our own wireless network under Arduino environment
to transmit data from one or more sensors at any time of the
day. Along with weather forecast information for
developing an algorithm for prediction of soil moisture of
the upcoming days. The system was first tested in our
laboratory and we plan to test it in a farm of vine near to El-
Malah city, located in western Algeria. Figure 3 shows
schematic of ML sub-process blocks for the execution of the
models. An operational architecture has been developed (see
Fig. 2) within the level of EDGE to predict the soil moisture
based on field sensors data, and also weather forecasting data
using LSTM-based models and GRU-based models.
In this paper, two different sources of information are
considered, each of them featuring complementary and
characteristic features useful to design and test LSTM and
GRU approaches: Historical hourly weather data 2012-2017
[24]: the data have been collected from 30 US and Canadian
Cities, as well as 6 Palestine cities. The dataset contains ~5
years of high temporal resolution (hourly measurements) data
of various weather attributes, such as temperature, humidity
what concerns us of our study. Detailed hydro meteorological
data from the mountain rain-to-snow transition zone are
present for water years 2004 through 2014 [23]: we interested
only to the soil moisture measurements. This version of the
data set fixes errors in all data files and supersedes the earlier
datasets. In this study, the LSTM cell architecture described
by Zaremba et al. [22] implemented in TensorFlow 2.0 has
been used for the experiments. Many factors may contribute to
speeding up the training time, including the dataset size, the
development platform (e.g., Tensorflow, Pytorch, Keras,
Caffe, MXNet), the hard-ware platform (e.g., CPU, GPU,
TPU) and the AI/ML model’s hyper-parameters (e.g., the
number of hidden layers, the number of neurons in each layer,
the learning rate, the batch size, the number of epochs).
Google Colab (GC) is an awesome place provided by Google
for training models. We was training our model but the Google
Colab keeps disconnecting after 2 hours automatically if we do
not respond the data is lost.
In time of prevent GC from disconnecting that's why in all
the experiments, we use a Laptop (Laptop PC) with processor:
Authorized licensed use limited to: Cornell University Library. Downloaded on May 24,2021 at 15:31:06 UTC from IEEE Xplore. Restrictions apply.
Intel Core i5-6300U @ 2.40GHz, Memory: 8GO RAM, Disk:
256 GO SSD.
B. Discussion and Results
In addition, the comparison between the performances of
the LSTM-based models and GRU-based models over the two
test sets together is depicted in figures 5, 6 and 7, by
comparing results on both training and validation datasets,
respectively. The analysis of the obtained results has revealed
the following key insights: (1) the training time can be
shortened by reducing the model size and increasing the batch
size; (2) CPU outperforms GPU in speeding-up the training
time of small-sized models. This is explained by the fact that
with small models, the CPU-GPU data transfer overhead
exceeds the computation acceleration benefit; (3) LSTM-
based models exhibit long training time compared to GRU-
based models.
Even LSTM (Long Short-Term Memory Recurrent
Networks) and GRU involve 1 input layer and 2 hidden layers
with 128 neurons for the first and 64 neurons for the second.
From the plot of loss, we can see that the model has
comparable performance on both training and validation
datasets. If these parallel plots start to depart consistently, it
might be a sign to stop training at an earlier epoch (See figures
5, 6 and 7). We discovered also the importance of collecting
and reviewing metrics during the training of our deep learning
models. Indeed, we observe that the training time for each
epoch with LSTM-based models is 200 sec, however it is
about 300 sec with GRU-based models. According to the
recent research results, it even surpasses LSTM-based models
in many applications. The experiments were conducted for the
prediction of the (Air-Temperature, Air-Humidity and Soil-
Moisture) to predict these parameters of the upcoming days.
Figures 5(a) and 5(b) illustrate that the model has comparable
performance for Air-temperature and gives best results on
both training and validation especially with LSTM-based
models even the Mean Squared Error (MSE) is higher
compared to GRU-based models. Figures 6(a) and 6(b) depict
that the model with GRU-based models outperforms slightly
LSTM-based models. Figures 7(a) and 7(b) show that the
parallel plots start to depart consistently after 10 epochs
approximately at the start of 60th epoch, it might be a sign
to stop training at an earlier epoch.
Despite all this, the model well mannered on both training
and validation datasets knowing that the scale of the graph is
too small. To avoid local minima problems in neural networks
using ReLU as the inner activation function, the experiments
are repeated 5 times and the median Mean Absolute Error
(MAE), the (MSE) and Root Mean Squared Error (RMSE) on
both test sets are presented in Table 1. According to the
results, GRU-based models can result better performance for
Soil-Moisture and Air-Temperature predictions. However,
LSTM-based model outperforms GRU-based models for Air-
Humidity predictions.
 (a) (b)
Fig. 5: (a) Plot of Model Loss on Training and Validation of Air-Temperature using LSTM-based models. (b) Plot of Model Loss on Training and Validation
of Air-Temperature using GRU-based models.

(a) (b)
Fig. 6: (a) Plot of Model Loss on Training and Validation of Soil-Moisture using LSTM-based models. (b) Plot of Model Loss on Training and Validation
of Soil-Moisture using GRU-based models.

(a) (b)
Fig. 7: (a) Plot of Model Loss on Training and Validation of Air-Humidity using LSTM-based models. (b) Plot of Model Loss on Training and Validation
of Air-Humidity using GRU-based models.

Table. 1: RMSE, MSE, MAE of the LSTM and GRU based models for (Air-Temperature, Soil-Moisture, Air-Humidity) Training and Validation.

Parameters LSTM GRU

RMSE MSE MAE RMSE MSE MAE

Soil-Moisture 0.0268 0,00072 0.0114 0.0220 0,00048 0.0101

Air-Humidity 11.7550 138,18 8.1946 12.9778 168.4255 8.4585

Air-Temperature 1.4090 1.9855 1.0777 1.3331 1.7771 1.0167

Authorized licensed use limited to: Cornell University Library. Downloaded on May 24,2021 at 15:31:06 UTC from IEEE Xplore. Restrictions apply.

V. CONCLUSION AND PERSPECTIVES
The implemented smart farming system was found to be
feasible and cost effective for optimizing water resources for
agriculture of precision. The main objective of this paper was
to design a new EDGE-Fog-IoT-Cloud based architecture
dedicated to the smart farming. We showed that AI techniques
play a pivotal role in agriculture of precision by using machine
learning and open sources technologies. One of future research
directions related to experiments is to collect the physical
parameters of our own farming system in order to collect our
dataset and the use of these sensors data along with weather
forecast information for developing an algorithm for
prediction of soil moisture of the upcoming days. There is also
the need to measure the hardware performance at the server
level when ingesting data. Meanwhile, we spotlighted the
limitations of AI techniques especially in training speed and
accuracy balance that may hamper the integration of ML.
Thus, more research efforts in this direction are required to
fulfil the potential of transfer learning.
ACKNOWLEDGMENT
This work has been supported scientifically by RIIR
laboratory/LAPECI laboratory. This paper is supported by the
PRIMA programme under grant agreement No 1821, project
WATERMED4.0. The PRIMA programme is supported by the
European Union.
REFERENCES
[1] J. Gutiérrez, J. F. Villa-Medina, A. Nieto-Garibay and M. Á. Porta-
Gándara, ‘Automated Irrigation System Based on Wireless Sensor
Network and GPRS Module’, vol. 63, no. 1, pp. 166–176, 2014.
[2] I. F. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci, “Wireless
sensor networks: a survey,” Computer Networks ,vol.38 , No.4,
pp.393–422, 2002.
[3] X. Wang, W. Yang, A. Wheaton, N. Cooley, and B. Moran, Efficient
registration of optical and IR images for automatic plant water stress
assessment, Comput. Electron. Agric., 2010.
[4] K. S. Nemali and M. W. van Iersel, An automated system for controlling
drought stress and irrigation in potted plants, Sci. Hortic.
(Amsterdam)., 2006.
[5] A. Goap, D. Sharma, A. K. Shukla, and C. Rama Krishna, An IoT based
smart irrigation management system using Machine learning and
open source technologies, Comput. Electron. Agric., vol. 155, no.
May, pp. 41–49, 2018.
[6] A. Dahane and N. E. Berrached, Mobile wireless and sensor, A clustering
algorithm for energy efficiency and safety, Apple Academic Press,
ISBN 9781771886796, July 2019.
[7] L. García, L. Parra, J. M. Jimenez, J. Lioret and P. Lorenz , IoT-Based
Smart Irrigation Systems: An Overview on the Recent Trends on
Sensors and IoT Systems for Irrigation in Precision
Agriculture, Sensors,Vol.20, No.4, pp.1042, 2020.
[8] M. C. Vuran, A. Salam, R. Wong, and S. Irmak, ‘Internet of underground
things in precision agriculture: Architecture and technology
aspects’, Ad Hoc Networks, vol. 81, pp. 160–173, 2018.
[9] S. Hwang, Monitoring and Controlling System for an IoT Based Smart
Home, Int. J. Control Autom., vol. 10, no. 2, pp. 339–348, 2017.
[10] M. Chakrabor ty, L. R. Khot and R. T. Pete rs, Assessing suitability of
modified center pivot irrigation systems in corn production using
low altitud e aerial imaging techniques, Information Processing In
Agriculture, In press, 2019.
Authorized licensed use limited to: Cornell University Library. Downloaded on May 24,2021 at 15:31:06 UTC from IEEE Xplore. Restrictions apply.
[11] A. Bonfante, et al. LCIS DSS—an irrigation supporting system for water
use efficiency improvement in precision agriculture: A maize case
study, Network Protocols & Algorithms vol.176, 102646. 2019.
[12] M. Mekki, O. Abdallah, M. B. M. Amin, M. Eltayeb, T. Abdalfatah, and
A. Babiker, Greenhouse monitoring and control system based on
wireless Sensor Network, Proc. - 2015 Int. Conf. Comput. Control.
Networking, Electron. Embed. Syst. Eng. ICCNEEE 2015, pp. 384–
387, 2016.
[13] A. Patil, M. Beldar, A. Naik, and S. Deshpande, Smart farming using
Arduino and data mining, Proc. 10th INDIACom; 2016 3rd Int.
Conf. Comput. Sustain. Glob. Dev. INDIACom 2016, pp. 1913–
1917, 2016.
[14] X. Wang and Q. Qi, Design and realization of precision agriculture
information system based on 5S, Proc. - 2011 19th Int. Conf.
Geoinformatics, Geoinformatics 2011, pp. 2–5, 2011.
[15] S. R. Prathibha, A. Hongal, and M. P. Jyothi, IoT Based Monitoring
System in Smart Agriculture, Proc. - 2017 Int. Conf. Recent Adv.
Electron. Commun. Technol. ICRAECT 2017, pp. 81–84, 2017.
[16] H. S. Kumar and S. Kusuma, Automated Irrigation System Based on
Wireless Sensor Network and GPRS Module, Int. Res. J. Eng.
Technol., vol. 3, no. 4, pp. 148–151, 2016.
[17] P. Singh and S. Saikia, Arduino-based smart irrigation using water flow
sensor, soil moisture sensor, temperature sensor and ESP8266 WiFi
module, IEEE Reg. 10 Humanit. Technol. Conf. 2016, R10-HTC
2016 - Proc., 2017.
[18] M. Rahman, Precision agriculture in the World and its prospect in
Bangladesh Research in Agriculture , Livestock and FISHERIES,
An open access peer Rev. J., vol. 3, no. April, pp. 1–14, 2016.
[19] L Parra, J Rocher, L García, et al. Design of a WSN for smart irrigation
in citrus plots with fault-tolerance and energy-saving algorithms,
Network Protocols & Algorithms vol.10, no. 2, pp. 95-115. 2018.
[20] A. Dahane, B. Kechar, ‘Remote Sensing Data: Automated Irrigation
Management Platform Using a WSNs’,International Conference on
Embedded Systems in Telecommunications and Instrumentation
(ICESTI'19), Annaba, Algeria, October, 28-30, 2019.
[21] A. Dahane, B. Kechar, Y. Meddah and O. Benabdellah, Automated
Irrigation Management Platform using a Wireless Sensor
Network, 2019 Sixth International Conference on Internet of
Things: Systems, Management and Security (IOTSMS), Granada,
Spain, pp. 610-615, 2019.
[22] W. Zaremba, I. Sutskever and O.Vinyals, Recurrent neural network
regularization. Preprint arXiv: 1409.2329, 2014.
[23] C.L. Enslin, S. E. Godsey, D.Marks,P.R, Kormos,M.S.Seyfried, J. P.
McNamara, and T. E.Link,: Hydro meteorological observations from
the rain-to-snow transition zone: a dataset from the Johnston Draw
catchment, Reynolds Creek Experimental Watershed, Idaho, USA
V1.1, Earth Syst. Sci. Data Discuss., [Online]. Available:
https://doi.org/10.5194/essd-2016-44, 2016.
[24] Historical Hourly Weather Data 2012-2017, [Online]. Available:
https://www.kaggle.com/selfishgene/historical-hourly-weather-
data/metadata (accessed: 27 April, 2020) .